Building A Model - Search Criterion

Searching Random Model Space

model_distance <- function( interscept, slope, X, Y ) {
  yhat <- interscept + slope * X
  diff <- Y - yhat
  return( sqrt( mean( diff ^ 2 ) ) )
}

Searching Random Model Space

models$dist <- NA
for( i in 1:nrow(models) ) {
  models$dist[i] <- model_distance( models$beta0[i],
                                    models$beta1[i],
                                    df$x,
                                    df$y )
}
head( models )

      beta0       beta1     dist
1 13.350016  0.39903106 18.16574
2 28.442296 -3.28888908 27.79515
3 34.304693 -0.56181867 10.56044
4 -2.537737  1.28467730 27.95880
5 18.242664 -0.71961230 20.58538
6  3.842021  0.05816131 29.00487

The Best Coefficients

ggplot( models, aes(x = beta0, y = beta1, color = -dist)) + 
  geom_point( data = filter( models, rank(dist) <= 10), 
              color = "red",  size = 4) + 
  geom_point()

Specifying a Formula

For R to perform these analyses for us, we need to be able to specify the function that we will use to represent the data (just like we did for plot(y~x)). Assuming the predictor is named x and the response is y, we have:

y ~ x

y ~ x1 + x2 

y ~ x1 + x2 + x1*x2

Fitting A Model

The easiest model.

fit <- lm( y ~ x, data = df )
fit 

Call:
lm(formula = y ~ x, data = df)

Coefficients:
(Intercept)            x  
     17.280        2.625  

Model Summaries

Summaries of the overal model components.

summary( fit )

Call:
lm(formula = y ~ x, data = df)

Residuals:
    Min      1Q  Median      3Q     Max 
-7.9836 -4.0182 -0.8709  5.3064  6.9909 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)   
(Intercept)   17.280      4.002   4.318  0.00255 **
x              2.626      0.645   4.070  0.00358 **
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 5.859 on 8 degrees of freedom
Multiple R-squared:  0.6744,    Adjusted R-squared:  0.6337 
F-statistic: 16.57 on 1 and 8 DF,  p-value: 0.003581

Components within the Summary Object

Like cor.test(), the object returned from lm() and its summary object have several internal components that you may use.

names( fit )

 [1] "coefficients"  "residuals"     "effects"       "rank"         
 [5] "fitted.values" "assign"        "qr"            "df.residual"  
 [9] "xlevels"       "call"          "terms"         "model"        

names( summary( fit ) )

 [1] "call"          "terms"         "residuals"     "coefficients" 
 [5] "aliased"       "sigma"         "df"            "r.squared"    
 [9] "adj.r.squared" "fstatistic"    "cov.unscaled" 

Components within the Summary Object

The probability can be found by looking at the data in the F-Statistic and then asking the F-distribution for the probability associated with the value of the test statistic and the degrees of freedom for both the model and the residuals.

summary( fit )$fstatistic 

   value    numdf    dendf 
16.56838  1.00000  8.00000 

The Summary Object

When you print out the summary( fit ) object, it uses the fstatistic to estimate a P-value. If we need that P-value directly, this is how it is done.

get_pval <- function( model ) {
  f <- summary( model )$fstatistic[1]
  df1 <- summary( model )$fstatistic[2]
  df2 <- summary( model )$fstatistic[3]
  p <- as.numeric( 1.0 - pf( f, df1, df2 ) )
  return( p  )
}

get_pval( fit )

[1] 0.0035813

Model Diagnostics - Residuals

plot( fit, which = 1 )

Normality Of the Data

plot( fit, which = 2 )

Leverage

plot( fit, which=5 )

Decomposition of Variance

The terms in this table are:

• Degrees of Freedom (df): representing 1 degree of freedom for the model, and N-1 for the residuals.

Decomposition of Variance

The terms in this table are:

• Sums of Squared Deviations:
  • \(SS_{Total} = \sum_{i=1}^N (y_i - \bar{y})^2\)
  • \(SS_{Model} = \sum_{i=1}^N (\hat{y}_i - \bar{y})^2\), and
  • \(SS_{Residual} = SS_{Total} - SS_{Model}\)

Decomposition of Variance

The terms in this table are:

• Mean Squares (Standardization of the Sums of Squares for the degrees of freedom)
  • \(MS_{Model} = \frac{SS_{Model}}{df_{Model}}\)
  • \(MS_{Residual} = \frac{SS_{Residual}}{df_{Residual}}\)

Decomposition of Variance

The terms in this table are:

• The \(F\)-statistic is from a known distribution and is defined by the ratio of Mean Squared values.

• Pr(>F) is the probability associated the value of the \(F\)-statistic and is dependent upon the degrees of freedom for the model and residuals.

Decomposition of Variance

The classis ANOVA Table

anova( fit )

Analysis of Variance Table

Response: y
          Df Sum Sq Mean Sq F value   Pr(>F)   
x          1 568.67  568.67  16.568 0.003581 **
Residuals  8 274.58   34.32                    
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Variance Exaplained

summary( fit )

Call:
lm(formula = y ~ x, data = df)

Residuals:
    Min      1Q  Median      3Q     Max 
-7.9836 -4.0182 -0.8709  5.3064  6.9909 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)   
(Intercept)   17.280      4.002   4.318  0.00255 **
x              2.626      0.645   4.070  0.00358 **
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 5.859 on 8 degrees of freedom
Multiple R-squared:  0.6744,    Adjusted R-squared:  0.6337 
F-statistic: 16.57 on 1 and 8 DF,  p-value: 0.003581

Relationship Between \(R^2\) & \(r\)

How much of the variation is explained?

\[ R^2 = \frac{SS_{Model}}{SS_{Total}} \]

c( `Regression R^2` = summary( fit )$r.squared,
   `Squared Correlation` = as.numeric( cor.test( df$x, df$y )$estimate^2 ) )

     Regression R^2 Squared Correlation 
          0.6743782           0.6743782 

Helper Functions

Grabbing the predicted values \(\hat{y}\) from the model.

df$yhat <- predict( fit )
df %>% 
  ggplot( aes(x,yhat) ) + geom_line() + geom_point()

Helper Functions - Residuals

The residuals are the distances between the observed value and its corresponding value on the fitted line.

df$residuals <- residuals( fit ) 
df %>%
  ggplot( aes(yhat, residuals) ) + 
  geom_point() + 
  geom_abline( slope=0,intercept=0, color="red")

What Makes One Model Better

There are two parameters that we have already looked at that may help. These are:

• The \(P-value\): Models with smaller probabilities could be considered more informative.

• The \(R^2\): Models that explain more of the variation may be considered more informative.

New Data Set

Let’s start by looking at some air quality data, that is built into R as an example data set.

airquality %>%
  select( -Month, -Day ) -> df.air
summary( df.air )

     Ozone           Solar.R           Wind             Temp      
 Min.   :  1.00   Min.   :  7.0   Min.   : 1.700   Min.   :56.00  
 1st Qu.: 18.00   1st Qu.:115.8   1st Qu.: 7.400   1st Qu.:72.00  
 Median : 31.50   Median :205.0   Median : 9.700   Median :79.00  
 Mean   : 42.13   Mean   :185.9   Mean   : 9.958   Mean   :77.88  
 3rd Qu.: 63.25   3rd Qu.:258.8   3rd Qu.:11.500   3rd Qu.:85.00  
 Max.   :168.00   Max.   :334.0   Max.   :20.700   Max.   :97.00  
 NA's   :37       NA's   :7                                       

Base Models - What Influences Ozone

Individually, we can estimate a set of first-order models.

\[ y = \beta_0 + \beta_1 x_1 + \epsilon \]

fit.solar <- lm( Ozone ~ Solar.R, data = df.air )
fit.temp <- lm( Ozone ~ Temp, data = df.air )
fit.wind <- lm( Ozone ~ Wind, data = df.air )

More Complicated Models

Multiple Regression Model - Including more than one predictors.

\(y = \beta_0 + \beta_1 x_1 + \beta_2 x_2 + \epsilon\)

fit.temp.wind <- lm( Ozone ~ Temp + Wind, data = df.air )
fit.temp.solar <- lm( Ozone ~ Temp + Solar.R, data = df.air )
fit.wind.solar <- lm( Ozone ~ Wind + Solar.R, data = df.air )

More Complicated Models

For Completeness

How about all the predictors. \(y = \beta_0 + \beta_1 x_1 + \beta_2 x_2 + \beta_3 x_3 + \epsilon\)

fit.all <- lm( Ozone ~ Solar.R + Temp + Wind, data = df.air )

For Completeness

\(R^2\) Inflation

Any variable added to a model will be able to generate Sums of Squares (even if it is a small amount). So, adding variables may artifically inflate the Model Sums of Squares.

Example:

What happens if I add just random data to the regression models? How does \(R^2\) change?

Random Data Effects

Random Data Effects

Perfect - My Models RULE

I can just add random variables to my model and always get an awesome fit!

Not so fast Bevis!

Model Comparisons

Akaike Information Criterion (AIC) is a measurement that allows us to compare models while penalizing for adding new parameters.

\(AIC = -2 \ln L + 2p\)

The criterion here are to find models with the lowest AIC values.

Questions